Builds the "indexed library" foundation and first intelligent features on
top of the organizer (sql.js index, non-destructive in-place indexing).
Phase 0 — Library index:
- sql.js (WASM SQLite) index DB; contentHash-keyed assets, resumable indexing
(skip by path+mtime), batch persistence (chosen over native better-sqlite3
which fails to build on Node 24 / Python 3.12)
- Library folders (in place, non-destructive) + background indexer w/ progress
- Thumbnails generated in the AI worker (canvas->webp), cached in userData;
served via photoai-media://thumb by hash; thumbnail grid w/ pagination
Phase 1 — AI quality assessment & culling:
- Focus (Laplacian variance), exposure (histogram), eyes-open (face-api EAR)
computed in one analyze pass alongside the thumbnail
- Culling filters (candidate/rejected) + quality badges
- Adjustable thresholds (live SQL re-classification from stored raw scores,
no re-analysis) + manual star rating (0-5) and color labels (usermeta)
Phase 2 — CLIP natural-language / similarity search:
- @huggingface/transformers (WASM/WebGPU, no native build)
- CLIP image/text embeddings (lazy-loaded); Korean queries auto-translated
via opus-mt-ko-en into the English CLIP
- Embeddings stored as SQLite BLOBs; "build search index" batch w/ progress;
brute-force cosine search; new Search tab
- Note: models download from HF Hub on first use; fully-offline ORT-wasm
packaging and KO search-accuracy tuning are follow-ups
Tabs added (Organize / Library / Search). All typecheck/tests(12)/build green;
boot smoke verified across phases.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
